Start Lesson
Here is what happens when you ask AI to do research without structure:
Research AI adoption in e-commerce.
You get 800 words of general background that reads like the first page of a Google search. "AI is transforming the e-commerce landscape..." followed by obvious trends anyone in the industry already knows. No sources. No confidence levels. No connection to your actual decision.
Now the same task with structure:
Act as a market research analyst specializing in e-commerce
technology adoption.
Research AI adoption in mid-market e-commerce companies
(50-500 employees) for a Shopify Plus agency deciding
whether to add AI implementation services.
For each finding, provide:
- CLAIM: The specific trend or data point
- SOURCE TYPE: Industry report, survey data, news, or
general knowledge
- CONFIDENCE: High / Medium / Low
- SO WHAT: What this means for our specific decision
Limit to 5 findings. Prioritize insights that directly
affect the build-or-wait decision. Skip general background.
You get a table of five specific findings, each with a confidence rating and a direct implication for your business decision. The "general knowledge" flags tell you which claims to verify. The "so what" column makes the output immediately usable in a strategy meeting.
That is the difference between research the AI wrote for no one and research the AI wrote for you.
After this lesson, you will be able to: prompt AI for structured research, competitor analysis, and devil's advocate stress-testing -- using CGC, roles, and chain-of-thought together.
This is the research template you will use most often. It combines role prompting (Lesson 2) with chain-of-thought structure (Lesson 3) to produce findings you can actually trust and use.
Act as a [role] with expertise in [domain].
Research [specific topic] for a [your role/company type]
who needs to decide [the specific decision this research
supports].
For each finding, provide:
- CLAIM: The key insight or data point
- SOURCE TYPE: Industry report / academic research / survey
data / news / general knowledge
- CONFIDENCE: High (well-established, multiple sources) /
Medium (credible but limited data) / Low (anecdotal or
inference)
- IMPLICATION: What this means for [your specific situation]
Limit to [number] findings. Prioritize actionable insights
over general background. If a claim is low-confidence, say
what I would need to verify it.
Expected output: A numbered list of findings, each with four labeled fields. The confidence ratings are the critical feature -- they tell you which findings to trust and which to verify before building a strategy around them. The "what to verify" instruction for low-confidence claims gives you a research to-do list, not just a report.
Works with Claude, GPT-4, and Gemini. Claude and GPT-4 tend to be more conservative with confidence ratings, which is what you want for business decisions.
Act as a competitive intelligence analyst.
Compare [Competitor 1], [Competitor 2], and [Competitor 3]
as alternatives to [your product/service].
For each competitor, analyze:
1. POSITIONING: Who they target and how they describe
themselves (use their actual language if possible)
2. STRENGTHS: What they do better than us -- be honest
3. WEAKNESSES: Where they fall short
4. PRICING: Their model and approximate price points
5. KEY DIFFERENTIATOR: The one thing that makes them
different from us
End with:
- A summary comparison table with columns: Company,
Target Customer, Price Range, Key Strength, Key Weakness
- One paragraph: where we have the clearest competitive
advantage and where we are most vulnerable
Be direct. I need honest analysis, not a document that
makes us feel good.
Expected output: A structured breakdown of each competitor followed by a comparison table you can paste into a deck. The "be honest" and "not a document that makes us feel good" constraints prevent the AI from defaulting to flattering your position -- a common failure mode in competitive analysis prompts.
This is the most valuable research pattern in this course. Instead of asking the AI to validate your plan, ask it to destroy your plan. Then ask it to synthesize.
Step 1 -- Attack:
Act as a skeptical board member who has seen many similar
plans fail.
I am planning to [describe your plan in 2-3 sentences,
including the investment and expected outcome].
Argue against this plan. Give me:
1. The 5 strongest reasons this could fail
2. The assumptions I am making that might be wrong
3. What evidence I would need to see to justify this
investment
4. The most likely way this fails even if the idea is sound
(execution risk)
Do not soften the criticism. I need the strongest possible
counterarguments.
Step 2 -- Synthesize:
Now evaluate both sides -- my original plan and your
counterarguments.
Which of your concerns are most legitimate?
Which ones can be mitigated, and how?
What are the 2-3 things I should investigate or test
before committing?
Give me a final recommendation: proceed, proceed with
modifications, or stop. Justify it in 3 sentences.
Expected output from Step 1: Five specific, uncomfortable objections -- not generic "it might not work" but pointed concerns like "your customer acquisition cost assumes a 4% conversion rate on cold outreach, but industry benchmarks for this category are 1.2%." Expected output from Step 2: A balanced assessment that acknowledges which objections are real threats vs. manageable risks, plus a clear recommendation with specific pre-conditions.
This two-step pattern produces more rigorous thinking than any single prompt. Use it before committing budget, launching a product, or making a hiring decision.
AI research has real boundaries. Ignore these and you will build strategy on fiction.
It fabricates sources. AI will cite papers, statistics, and quotes that do not exist. If a finding drives a business decision, verify it independently. The confidence ratings in Template 1 help -- treat anything below "High" as a hypothesis, not a fact.
It has a knowledge cutoff. Unless connected to a live search tool, it cannot tell you what happened last month. Ask it to flag any claims that may be outdated.
It defaults to the mainstream view. AI research skews toward well-documented, English-language sources. Niche markets, emerging regions, and contrarian positions may be underrepresented.
The mental model: AI is a research assistant, not a research authority. It drafts the brief. You verify the facts that matter.
Pick a real business decision you are facing -- a tool to buy, a market to enter, a feature to build. Run the Devil's Advocate pattern (Template 3) with both steps.
Step 1: Describe your plan and ask for the five strongest counterarguments.
Step 2: Ask it to synthesize and give you a proceed / modify / stop recommendation.
Check your output: The counterarguments should make you uncomfortable. If they are all softballs ("there is some risk involved"), your plan description was too vague -- add specific numbers, timelines, and expected outcomes so the AI has real assumptions to challenge.
You now have templates for writing (Lesson 4) and research (this lesson). In the final lesson, you will learn how to systematize your prompts -- building a reusable library with templates, variables, and naming conventions so you never write the same prompt from scratch twice.